Inference with "difference in differences" with a small number of policy changes